We were seeing a drop in engagement on our core feature and leadership wanted answers fast. We didn't have clean data, so I pulled what we had, ran a quick analysis, and identified that the issue was likely tied to a recent onboarding change. I worked with engineering to roll back that change and engagement recovered within two weeks. It taught me to move quickly when the data is messy rather than waiting for perfect information.
When retention dropped seven percent in a core cohort, there was no clean signal — three teams had shipped changes the same week. Instead of chasing the most obvious suspect, I wrote down every hypothesis, ranked them by testability and user impact, and asked what data we'd need to falsify each one. That process took two days and surfaced a hypothesis no one had considered: a silent permission change affecting low-bandwidth users. We ran a targeted experiment, confirmed causality, and recovered the cohort within three weeks. The structured framing is what got us to the right problem, not the right answer first.